Refactorization of the EVP solver

Earth system models (ESM) strive to describe more and more details. This is often accomplished with the use of more complex descriptions or higher resolution. The limitation of this approach is often the computer system at hand. In many cases, ESM’s are written with a focus on the physical system development and less on how to structure the code according to infrastructure on the computer. This presentation focuses on the sea ice dynamics, and particularly on the solver for the Elastic-Viscous-Plastic (EVP) equations. The EVP approach introduces artificial elastic waves that are iteratively dampened. Hundreds of iterations are necessary to reach a solution. In the traditional implementation, each iteration requires communication between the processors using MPI calls.

An analysis of the existing solver’s performance was first carried out based on the sea ice model CICE. Three performance challenges were identified with the current implementation: Two challenges relate to the parallelization itself, namely 1) General imbalance issues due to the nature of the challenge and 2) MPI synchronization after each sub-cycling. The third issue relates to the data structures chosen and their corresponding memory access patterns. This study aim at removing all three limiting factors by adjusting the memory access patterns and by adjusting the parallelization approach so that we can avoid the costly MPI synchronization after each sub-cycle, which enables the use of parallel instructions, which are available in modern hardware. The adjusted implementation runs significantly faster.

EVP is just one component out of many others in sea ice/ESM model (and other modelling systems). The refactoring includes a novel integration that shows how the EVP solver can be integrated into various model systems via the MPMD pattern and hence also runs on heterogeneous systems.